Style Guides Don't Work for AI—Profiles Do
By Jack Stovell X · Instagram · LinkedIn · 2026-10-26 · General
Your style guide is fine. It just isn't doing what you think it's doing.
Here's the problem. A human reads "confident but approachable" and pulls on years of context: who the company is, what the last launch email sounded like, what the CEO would wince at. A model reads the same two adjectives and does arithmetic. It averages every confident-and-approachable text it has ever seen and hands you the middle of the road. That's why drafts from five contributors, all pasting in the same guide, still sound like five different people. Or worse, like nobody.
So the fix isn't a better guide. It's a different kind of input.
What a style guide looks like to a model
A prose style guide is a pile of descriptions. "Conversational." "Punchy." "Avoid jargon." Each one is a label pointing at a behaviour, and the model has to guess the behaviour from the label. Guesses vary. Change one word in the prompt, add a long source document, switch the task from a blog post to a product email, and the weighting shifts. The instruction is still technically present. It just matters less than it did.
And to be fair, the model isn't being lazy. Prose instructions compete with everything else in the context window. Your tone note sits next to the topic brief, the audience description, the examples, the formatting rules. Attention is finite. Adjectives lose.
Where does it hallucinate style? Mostly in the gaps. "Punchy" doesn't say how short. "Direct" doesn't say whether that means no hedging or just no preamble. The model fills the gap with a stereotype of punchy, which usually means exclamation marks and sentence fragments piled on like confetti. Fair enough as a reading. Wrong as a match.
What a stylometric profile looks like to a model
A stylometric profile replaces the label with the measurement. Average sentence length. How much that length varies. Contraction rate. How often the writer opens with the conclusion. Comma density, the ratio of "the" to "a," whether lists are numbered or bulleted, which filler phrases recur. It's a structured block of data, not a paragraph of encouragement.
The difference matters mechanically. "Short sentences" is a suggestion. A target average of roughly seventeen words with high variance is a constraint the model can check its own output against, token by token. There's nothing to interpret. The number is the number.
This is also why it survives prompt variation. A tone adjective is a semantic instruction, so its effect depends on everything around it. A structured profile is a pattern the model matches against, and patterns hold their shape when the surrounding prompt changes. Swap the task, change the topic, lengthen the brief. The rhythm targets still read the same way every time. You've moved the style signal out of the part of the prompt that gets reinterpreted and into the part that gets copied.
A worked example
Imagine a team writing a product update. The style guide says: "Our voice is direct, warm and a little dry. We lead with the point."
Prompted with only that, a model might produce:
"We're thrilled to share some exciting news about our latest update! Over the past few months, our team has been working hard to improve your experience, and we think you're going to love what's new."
Nobody on that team has ever written that sentence. It's warm in the way a greeting card is warm. It buries the point under two clauses of throat-clearing, and "thrilled" and "exciting" are exactly the stereotype filling the gap.
Now give the same model a profile: conclusion-first structure, zero exclamation marks, around seventeen words per sentence with wide variance, high contraction rate, short fragments after longer sentences, dry humour register. Same task, same topic:
"Reporting got faster. That's the headline. We rebuilt how dashboards load, so the page you open every morning no longer makes you wait. Fair enough if you didn't notice the old delay; you will now."
It leads with the point. It uses contractions. It drops a fragment after a longer sentence and lets the dryness do the warmth. Nobody told it to avoid "thrilled." The measurements made that word statistically foreign.
Moving your team over
You don't need to throw the style guide away. Treat it as the why, and the profile as the how. Here's the sequence.
- Gather your best writing. Pick pieces your team agrees sound like you, and keep them recent. Old samples drag the profile backwards.
- Measure it. Extract the quantified features from those samples: sentence rhythm, punctuation habits, structural patterns, recurring phrases. ScriptGrain does this extraction and produces a profile you can inject as structured context, which beats counting commas by hand.
- Decide what's brand and what's author. A company profile and an individual contributor's profile are different objects. Layer them deliberately rather than blending them into mush.
- Inject, don't describe. Put the profile in as structured data at the top of every prompt, not paraphrased into a paragraph. Paraphrasing turns measurements back into adjectives, and you're right back where you started.
- Test against the reference. Run the same brief through five different prompt phrasings. If the rhythm holds across all five, the layer works. If it drifts, tighten the constraints.
- Review quarterly. Writing evolves. Re-measure on fresh samples and update.
The short version
Style guides describe a voice. Profiles specify one. Models are far better at matching a spec than interpreting a description, and that's the whole trick. Your contributors don't need to become better prompt writers; they just need a context layer that doesn't get lost in translation.
At the end of the day, you can keep asking a model to be "confident but approachable," or you can hand it the numbers. One of those has a track record of producing the same voice twice.